A Reーiabiーity Study 。f the G-TELP Test` Leveー 2

Transcription

A Reーiabiーity Study 。f the G-TELP Test` Leveー 2
AReliability Study of the G−TELP Test, Level 2
Laura MacGregor
Introduction
This paper seeks to investigate the reliability of the G−TELP test, Level
2,Form 211, which was used to select eligible students for the Intensive
English Course at Gakushuin University in April,2002. Of the 237 students
who took the test,46 were admitted to the programme based on a cutoff
score determined after the test.Using classical descriptive and item analy−
ses, the Gakushuin data are examined here to assess how the items are
performing, to determine the test’s reliability, and finally, to assess the
G−TELP’s suitability as a placement instrument.
Overview of the Test
General information
The G−TELP is a 5−level criterion−referenced English proficiency test
developed and maintained by San Diego State University. It is currently
being used by 71 institutions(high schools, junior colleges, universities)in
Japan, primarily as a placement instrument. Level 2, the subject of this
study, is an 80−item test consisting of three sections:grammar(26 items),
listening(26 items), and reading and vocabulary(28 items). Each of the
section scores are reported out of 100 for a total score of 300. To be ranked
at Level 2, test−takers must score 750r higher in all three sections.
According to the test distributors, the G−TELP, which tests“authentic
and modified English in normal communication,”is designed to“measure
language proficiency in two broad areas:functional ability and grammati−
cal ability”(G−TELP Japan office, personal communication, April,2002).
103
『言語・文化・社会」第1号
The general description for Level 2 states that:
This level assesses the ability of an examinee to use the language outside of
classroom situations. This examinee is able to cope with some authentic English and
has experienced contact with native speakers. Although his/her learning of the
language has been classroom−based, the examinee is able to communicate with a
native speaker within a wide range of tasks.(San Diego State University,1998).
Itemり1ρes
Since the actual test forms are reused by G−TELP and are therefore
confidential, the contents of Form 211 cannot be disclosed. However, a
study guide published by Kinseido(Morita,1998)gives the following
general information about the Level 2 test.
Level 2 addresses a finite number of skills, which are identified as
follows according to test section. First, the grammar section includes items
which address four grammar points:(i)participials(7 items);(ii)present
unreal(7 items);(iii)past perfect tense(6 items);and(iv)future progres−
sive tense(6 items). Four listening tasks address comprehension of:(i)a
narrative;(ii)the description of a process;and conversations involving(iii)
persuasion;and(iv)negotiation. The length of each listening passage in the
sample guidebook averaged 2 minutes 51seconds(ranging from 2 minutes
13seconds to 3 minutes 35 seconds each). The reading and vocabulary
section also has four sections which present four different text types in
passages of 300−400 words:(i)ahistorical account;(ii)an article on a
technical or social topic;(iii)an encyclopedia entry;and(iv)abusiness
letter with descriptive and persuasive elements. There are three question
types in this section:(i)reading comprehension(literal skills);(ii)inferen−
tial skills;and(iii)vocabulary knowledge.
104
AReliability Study of the G−TELP Test, Level 2(Laura MacGregor)
Method
The data were first analyzed to generate descriptive statistics illustrat−
ing features of central tendency and dispersion. To determine how well the
individual items performed on this administration of the test, the item
facility(IF)and item discrimination(ID)values were calculated for each
section. Finally, the reliability coefficient for the test was generated using
the Kuder−Richardson 20(K−R20)formula. The software used was Micro−
soft Excel for Mac(Microsoft,2001).
Results
(a) Descriptive statistics
Descriptions of the central tendency(mean and median)are reported
in Table l to identify the typical, or common scores for the test and for
each section. Dispersion figures, which indicate how individual scores
disperse around the central tendency are also reported(Table 1).
Table 1:Descriptive Statistics(N=237)
Dispersion
Central Tendency
k
Mean(%)
Low・High
Median(%)
Range
S
Grammar
26 60.9 62
15−100
86
15.417
15−100
86
17.250
11−100
90
16.959
61−281
221
39.255
Listening
26 57.1 58
Reading and Vocabulary
28 63.6 64
Total
80 181.6 183
105
『言語・文化・社会」第1号
,K!!gye
N=number of test−takers k=number of test items
Median=point below which 50%of the scores fall and above which 50%of the
scores fall
Range=high score−low score十1 S=standard deviation
The median scores are nearly the same as the mean scores, suggesting
anearly normal distribution. Central tendency alone however, does not
give a complete picture of distribution as it works together with dispersion.
Dispersion indicates how individual scores disperse around the central
tendency, and shows how widely the scores are spread out. The figures here
(Table 1)indicate that a normal distribution is not likely for the grammar
section or the reading and vocabulary section, and therefore for the whole
test, since the distance between the low scores and the mean is different
from the distance between the high scores and the mean. A look at the
histograms for the three sections and for the total test(Figures 1−4 in the
Appendix;score report data from G−TELP Japan office, April 8,2002)
confirms that no part of the test produces a normal distribution.
(b) ltem/iZcility and item discrimination
Next, item facility(IF)and item discrimination(ID)values are
reported. Item facility represents the percentage of test−takers who an−
swered the item correctly. Facility values are reported as factors of 1.00
and range from O.00 to 1.00. Item facility shows how easy or difficult the
test items are. An IF of.95 means that 95%of the test−takers got the item
right, a very easy item. If the purpose of the test is to produce a wide range
of scores, then items with IF values of around.50 should be used because
they offer the greatest potential for variation among test−takers(Alderson,
Clapham, and Wall,1985, p.81).Therefore, IF values can help control the
106
AReliability Study of the G−TELP Test, Level 2(Laura MacGregor)
difficulty level of the test.
Item discrimination shows how well an item distinguishes between test−
takers of different ability levels by examining the differences between the
high scorers and the low scorers. ID values are calculated by first ranking
test−takers according to their total scores, then comparing the proportion of
correct answers(IF)in the top(or upper)group with those in the bottom
(or lower)group. The formula for ID is:
ID=IF(upper)−IF(lower)
An ID of.400r higher is considered to be acceptable;however,“there
are no rules as to what[IDs]are acceptable, since the possibility of high
[IDs]varies according to the test type and range of ability of the exami−
nees” iAlderson, Clapham, and Wall,1995, p.82). Ebel(1979, quoted in
Brown,1996, p.70)gives the following guide for items with IDs below.40:
.30−.39
Reasonably good but possibly subject to improvement
.20−.29
Marginal items, usually needing improvement
Below.19
Poor items, to be rejected or improved by revision
It is important to consider the relationship between IF and ID. As
mentioned above, IFs close to.50, since they represent items of medium
difficulty, are effective for the purpose of spreading the test scores out over
awide range. For the same reason, the IFs should be as close to.50 as
possible to achieve high IDs, if the purpose of the test is to spread the test
scores out as far as possible. A hypothetical table of IFs which are possible
with certain ID values is presented as follows(Table 2, adapted from
Alderson, Clapham, and Wall,1995, p.84):
107
『言語・文化・社会」第1号
Table 2:Maximum IDs for IF Values
IF
1.0 .93 .80 .70 .66 .50 .33 .30 .20 .06 0.0
max ID
0.0 .20 .60 .90 1.0 1.0 1.0 .90 .60 .30 0.0
Tables 3−5 present the IF data for the total population of test−takers for
each of the three sections(using raw data provided by the G−TELP Tokyo
office, April,2002). Below the total IF figures, ID values are calculated
using the upper 25%and lower 25%of the sample populations. The ID
figures in bold, underlined type indicate items with ID values of below.19,
which will be addressed in the following discussion section. However, since
the actual contents of individual items are confidential, an examination of
the possible factors which caused these items to perform poorly(i.e., choice
of vocabulary or distractors, wording, and cultural appropriacy)cannot be
made.
Table 3:IF and ID for Grammar Section(N=26)
item#
1
2
3
4
5
6
7 8 9
total IF
0.78
0.79
0.82
0.46
0.57
0.48
0.84 0.41 0.57
IF upper 25%
0.78
0.94
0.87
0.70
0.76
0.58
0.90 0.43 0.70
IF lower 25%
0.68
0.71
0.77
0.24
0.42
0.37
0.75
0.39 0.54
ID
蜘
easy
0.23
蜘
easy
0.46
0.34
0.22
螂
easy
difficult easy
108
Q.Q4 0、ユ5
AReliability Study of the G−TELP Test, Level 2(Laura MacGregor)
item#
10 11
12 13
14 15
16 17
18
total IF
0.36 0.63
0.43 0.79
0.30 0.95
0.78 0.51
0.65
IF upper 25%
0.53 0.71
0.54 0.85
0.42 0.97
0.84 0.73
0.72
IF lower 25%
0.19 0.54
0.33 0.70
0.19 0.92
0.73 0.28
0.54
ID
0.34 Ω、ユ6
0.22 Ω、ユ5
0.23 Q.Qs
0、ユQ O.46
easy
easy
easy
easy
螂
easy
item#
19 20
21 22
23 24
25 26
total IF
0.52 0.57
0.55 0.76
0.62 0.38
0.71 0.62
IF upper 25%
0.65 0.77
0.67 0.85
0.80 0.58
0.90 0.72
IF lower 25%
0.37 0.39
0.32 0.61
0.43 0.18
0.54 0.52
ID
0.28 0.38
0.35 0.24
0.37 0.41
0.35 0.20
Table 4:IF and ID for Listening Section(N=26)
item#
27
28 29
30 31
32 33
34 35
total IF
0.51
0.59 0.35
0.54 0.59
0.56 0.43
0.75 0.49
IF upper 25%
0.67
0.75 0.62
0.70 0.76
0.78 0.66
0.85 0.73
IF lower 25%
0.41
0.41 0.16
0.38 0.41
0.37 0.23
0.57 0.32
ID
0.27
0.34 0.46
0.32 0.35
0.42 0.43
0.28 0.42
item#
36
37
38
39
40
41
42
43
44
0.87
0.49
0.80
0.73
0.50
0.75
0.64
0.49
0.57
0.92
0.85
0.71
0.95
0.78
0.51
0.75
0.41
0.35
鋤
0.39
IF upper 25%
IF lower 25%
0.96
0.77
IF lower 25%
0.77
0.23
0.66
0.62
0.37
0.66
0.41
ID
0.19
0.54
0.27
0.23
0.34
0.29
0.38
difficult
109
『言語・文化・社会」第1号
item#
45 46
47 48
49 50 51 52
total IF
0.34 0.34
0.46 0.56
0.69 0.39 0.85 0.57
IF upper 25%
0.52 0.47
0.63 0.77
0.81 0.44 0.92 0.76
IF lower 25%
0.24 0.29
0.28 0.33
0.52 0.29 0.75 0.38
ID
0.28 Ω、ユ8
0.35 0.44
0.29 Q.15 0.18 0.38
difficult
difficult easy
Table 5:IF and ID for Reading and Vocabulary Section(N=28)
60 61
62
0.89 0.54
0.72 0.64
0.73
0.96 0.85
0.96 0.70
0.86 0.78
0.87
0.67 0.39
0.80 0.68
0.76 0.42
0.56 0.41
0.56
0.20 0.33
Ωユ6Ωユ6 0.20 0.28
0.30 0.38
0.32
item#
53 54 55 56 57 58 59
total IF
0.93
0.79 0.53
0.89 0.77
IF upper 25% 0.97
0.87 0.72
IF lower 25%
0.82
ID
Q,15
easy
easy easy
item#
63
64 65
66 67 68 69 70 71
total IF
0.53
0.84 0.68
0.68 0.74
0.41 0.70
0.45 0.35
0.34
IF upper 25% 0.71
0.94 0.85
0.80 0.91
0.49 0.81
0.65 0.54
0.53
IF lower 25%
0.38
0.73 0.53
0.51 0.58
0.32 0.53
0.33 0.16
0.15
ID
0.33
0.20 0.32
0.29 0.33
Q.,ユ80.28
0.32 0.38
0.38
difficult
74
75
76
77
78
79
80
total IF
0.67
0.74
0.40
0.35
0.28
0.57
0.81
0.88
IF upper 25%
0.94
0.82
0.57
0.48
0.46
0.82
0.95
1.00
IF lower 25%
0.41
0.65
0.22
0.24
0.14
0.37
0.68
0.71
ID
0.53
螂
easy
0.35
0.24
0.32
0.46
0.27
0.29
item#
110
AReliability Study of the G−TELP Test, Level 2(Laura MacGregor)
ピc/Reliability
Reliability indicates how the scores on one test administration are
likely to be very similar to those which would be obtained if it had been
administered to the same students at another time. The greater the similar−
ity between these two sets of scores, the higher the reliability of the test.
While only one administration of the G−TELP occurred, the reliability can
still be estimated, using the Kuder−Richardson 20(K−R20)formula, the
most accurate formula for measuring reliability. The formula for K−R20 is
as follows:
k
K−R20=
k−1
ΣIV
(1−sl)
Key
ΣIV=sum of item variance for each item, where IV=IF(1−IF)
St2=variance for the whole test(standard deviation of the test scores squared)
The reliability of this form of G−TELP Level 2 was calculated to be
.85,meaning that there is an 85%chance that a second administration of
this test would produce very similar results. Unfortunately, there is no
standard acceptable level of reliability, and opinions vary regarding mini−
mum reliability coefficients according to the type of test. While there is no
documentation on what the reliability of a criterion−referenced test like the
G−TELP should be, Lado(quoted in Hughes,1989, p.20)suggested that
good vocabulary, structure, and reading tests are usually in the.90 to.99
「ange・
Discussion
As mentioned above, the item types are finite, and the grammar items,
in particular, are restricted to four grammar aspects which are repeated
111
『言語・文化・社会」第1号
over six or seven items each. It is not known how these four aspects were
determined or whether indeed they reflect the skills and abilities that a
successful Level 2 candidate should be able to demonstrate. Nor is it clear
how the types of listening and reading passages were selected and how they
contain discourse that reflects a Level 2 candidate.
The descriptive analyses of the test reveal that none of the sections
produced normal test score distributions. That there are two peaks in the
grammar and listening sections(Figures l and 2)indicates that these
sections are not functioning well. While a normal distribution is not
necessarily a requirement for a criterion referenced test, the distribution
histogram for the total test scores(Figure 4)indicates that the majority of
test−takers scored in the 150−210 range, clustering around the average of
181.6.This is irregular, since criterion referenced tests usually produce
either positively skewed curves(where the scores tail off in a smooth curve
towards the right end of the graph for an easy test)or negatively skewed
curves(where the scores tail off in a smooth curve towards the left end of
the graph for a difficult test. The fact that neither of those types of curves
exist here indicates that there are problems with the test.
It is important to keep in mind that with a criterion reference test, the
test−takers’scores are not being compared to each other(as in a norm
referenced test like TOEFL and TOEIC),but are interpreted purely based
on the number of items test−takers got right. The closer the score is to
100%(or at least 75%or higher for all three sections), the more closely
the test−taker fits the profile of proficiency that the test aims to measure.
Since the G−TELP was used at Gakushuin as a placement instrument to
determine which students should enter the Intensive English course, a test
which spread test−takers out over a normal distribution would have made it
clearer who the most proficient students(according to the parameters of
the test)were. If that is the case, then a norm referenced test, which
112
AReliability Study of the G−TELP Test, Level 2(Laura MacGregor)
addresses general skills of language proficiency and covers a wider range of
language and language use may be preferable. Further, the average scores
for the reading and vocabulary section were much higher than that of the
listening section(a 6.5%difference in the Gakushuin data and a 9%
difference in national data for 1,249 test−takers in Japan during the first
half of 2002), suggesting that test level is not uniform across sections, at
least for Japanese test−takers, and that the test may not be an accurate
measure of the skills it directly addresses. Further, since only 140f the 237
test−takers achieved a Level 2 score, the effectiveness of this test is
questionable.Therefore, from a decision−making perspective, the G−TELP
may not be the best instrument for placement purposes.
Item analyses revealed that 19(24%)of the 80 test items were very
weak(i.e., with ID values below.19), and therefore made no positive
contribution to the test.These items are listed in Table 6, according to
whether they were easy or difficult items(see also Tables 3−5).
Table 6 Summary of weak items with ID values below.19
Section
Easy items
Difficult items
Grammar
N=9(#1,#3,#7,#9,
N=1(#8)
Total(N)
10
#11,#13,#15,#16,#18)
N=1(#51)
N=3(#43,#36,#50)
Reading and
N=4(#53,#56,
N=1(#68)
Vocabulary
#57,#74)
4・5
Listening
Clearly, there are many items which are too easy for this group of
test−takers, suggesting that either the test is either entirely, or at least in
part too easy, or that the test simply contains too many weak items that
113
『言語・文化・社会」第1号
need to be replaced with items that discriminate better between high and
low scorers.
On the whole, the listening section discriminates most effectively, with
only four very weak items, but still needs improving. The weakest section
is the grammar section, which tests the following four grammar skills:
participials, present unreal, past perfect, and fUture progressive. Half of the
weak items in this section are participial items(N=5:#1,#8,#11,#13, and
#18).Therefore, participials are not really being tested here since five of
the seven items which test this point are extremely weak. On the other
hand, all of the future progressive items(#4,#6,#10,#14,#17,#24)are
functioning fairly well, and can be said to be good or at least acceptable
ltems.
The reliability coefficient of.85 is below Lado’s.90−.99 range and is
also below the.93 reliability figure of the TOEIC reading and listening
sections(Woodford,1991,p.12),likely because roughly one quarter of the
items are very weak. Therefore, the reliability of the G−TELP is question−
able.
Conclusion
The G−TELP test was used as a placement instrument to screen
applicants to the Intensive English Course at Gakushuin University. Using
descriptive analyses and item analyses, the test was examined to assess its
reliability. Descriptive analyses revealed that the none of the three sections
of the test were functioning normally since they did not produce normal
distributions. While a normal distribution is not necessarily a requirement
for a criterion referenced test, the clustering of scores around the mean
(Figure 4)and the appearance of double peaks in the grammar section and
listening section histograms(Figures l and 2)indicate that the test is not
functioning well, even for a criterion referenced test.
114
AReliability Study of the G−TELP Test, Level 2(Laura MacGregor)
The item analyses showed that 24%of the items were extremely weak,
and the majority of those items appeared in the grammar section. Thus, it
is difficult to know if the results of the test indeed reflect the general
description of what a Level 2 candidate can do.
The reliability coefficient of.85 is relatively low, suggesting that the
test may not be a strong enough instrument for the important task of
deciding who will be admitted to the Intensive English Course.
While there are a number of logistical factors to consider when
selecting a placement test−cost, ease of administration, time required to
generate and receive the results−the selection of an instrument which is
appropriate to the task and which produces accurate results is crucial.
Since there are obviously a number of problems with the G−TELP test, as
evidenced by this study, it may be advisable to choose a more appropriate
and reliable instrument for future test administrations.
Acknowledgements
1.Figures 1−4 are reproduced here with the permission of the G−TELP
Tokyo office, October 2002. Raw item data provided by G−TELP was also
used with permission.
2.Special thanks to Koarai Mikiya for his help with generating the
StatiStiCal reSUItS.
References
Alderson, J. C., Clapham, C.,&Wall, D.(1995). Language鰭
construction and evaluation. Cambridge:Cambridge University Press.
Brown, J. D.(1996).7セ3伽g in language program&Upper Saddle
River, NJ:Prentice Hall Regents.
Hughes, A.(1989). Testing/br language teachers. Cambridge:Cam一
115
『言語・文化・社会」第1号
bridge University Press.
Microsoft Corporation.(2001)..Microsoft.Eκcθ12001!mac(Japanese
version). Redmond, WA:Microsoft Corporation.
Morita, K.(1998).国際英検θ一7EL P.1,2. Tokyo:Kinseido.
San Diego State University.(1998, April 1).G−TELP Level Descriptors.
(booklet copied and distributed by the G−TELP Japan office)
Woodford, P.(1991, December 10). A historical overview of TOEIC
and its mission. In TOEIC Steering Committee(Ed.), The 35th TO.EIC
seminar (pp.10−15). Tokyo:The Institute for International Business
Communication.
116
AReliability Study of the G−TELP Test, Level 2(Laura MacGregor)
Appendix(Figures 1−4)
Figure 1:Grammar section score distribution
GRMヒZトグラム
80
1 OO.O%
60
80.o%
蟹4。
60.⑪%
臨2⑪
40.O%
o
o.⑪%
2⑪.⑪%
10 20 30 40
50 6⑪ 70 呂o go loo
データ区間(10点)
Figure 2:Listening section score distribution
100,0%
80,0%
60.o%
騒4⑪
40,0%
20.O%
o.o%
50 6e 70
データ区間(10点)
117
『言語・文化・社会」第1号
Figure 3:Reading and vocabulary section score distribution
ロ頻
一■一累積畷
RD(ヨヒストゲラム
80
19 ll
100。o%
9⑪.o%
7⑪
巽ll
80,0%
70,0%
6⑪.O%
5⑪,o%
40,0%
30,0%
2⑪.o%
1⑪,o%
⑪
o,o%
10 20 30 40
50 60 7⑪ {〕0 90 100
テ’一タ区間(10点)
Figure 4:Total test score distribution
目頻度
mtストグラム
一■一累積
80
1 OO.O%
一
8⑪.o%
60
60,0%
騒4⑪
4・0,0%
20
一一
2⑪.⑪%
o
o.〔隔
30 60 9⑪ 120 150 180 210 240 270 3⑪O
デー一タ区間(3e点}
118
G−TELP Test Level 2の信頼度に関する研究
ローラ マクレガー
本稿では、G−TELP Test Leve12の信頼度にっいて調査を行う。 G−TELP
Testとは、サンディエゴ州立大学が開発し、維持・管理を行っている基準参照
英語技能試験である。2002年4月、学習院大学インテンシヴ・コースの選抜
試験として、G−TELP Testが237名の学生に対して実施された。記述分析及
び項目分析を通して学習院大学の試験データを調査することによって、項目の
適切度及び試験としての信頼度の判定を行い、最後にプレイスメントテストと
しての適性にっいて結論を出したい。
記述分析では、どのセクションを見ても正規分布は認められなかったため、
G−TELP Testが正常に機能していないということが明らかとなった。項目分
析では、24%の項目(ほとんどが文法に関する項目)が極めて不適切であると
いうことが明らかとなった。.85という信頼度係数は、比較的低いものである。
したがって、インテンシヴ・コースの履修者を決定するという重要な作業を行
うためには、G−TELP Testは、不適切な手段であると言わざるを得ない。
プレイスメントテストを決定する際には、費用、手間、結果がでるまでにか
かる時間等の様々な要因が関わってくるが、正確な結果が得られる適切な試験
を選ぶことが肝腎である。本稿の研究で明らかなように、G−TELP Testには数
多くの問題点があるため、今後の試験には、より適切で、信頼性の高いものを
選ぶ必要があると思われる。
124一